Professional Documents
Culture Documents
OF VOTING
H. P. YOUNG
University of Maryland
[Our aim is] to inquire by mere reasoning, what own prevails, this proves only that I have
degree of confidence the judgment of assemblies made a mistake, and that what I believed
deserves, whether large or small, subject to a
high or low plurality, split into several different to be the general will was not so" (1962,
bodies or gathered into one only, composed of 153).
men more or less wise. This ambiguous and slightly disquieting
—Condorcet idea was given a more satisfactory expres-
sion some twenty years later by the math-
ematician and social philosopher Marie
central problem Jean Antoine Nicolas Caritat, Marquis de
in democratic theory is to justify the prin- Condorcet.1 In his remarkable work,
ciple of majority rule. Why should the Essai sur Vapplication de I'analyse a la
opinion of a majority of citizens or of probability des decisions rendues a la
elected representatives bind the rest of pluralite des voix (1785), Condorcet pro-
society? One of the earliest and most cele- posed the following argument for decision
brated answers to this question was at- by majority: Enlightened voters honestly
tempted by Rousseau in The Social Con- attempt to judge what decision will best
tract (1762). The opinion of the majority serve society. They may occasionally
is legitimate, said Rousseau, because it judge wrongly. But assuming that they
expresses the "general will." "When a law are more often right than wrong, the
is proposed in the people's assembly, majority opinion will very likely be "cor-
what is asked of them is not precisely rect." This is just a property of large num-
whether they approve of the proposition bers and is intuitively obvious when there
or reject it, but whether it is in conformity are only two possible decisions. Con-
with the general will . . . each by giving dorcet rigorously demonstrated this prop-
his vote gives his opinion on this ques- osition using the newly developed cal-
tion, and the counting of votes yields a culus of probabilities. He then proceeded
declaration of the general will. When, to show how the idea could be elaborated
therefore, the opinion contrary to my into a whole series of results about the
1232
Condorcet's Theory
pair of candidates and the results are tal- Table 1. Voting Matrix with
lied in a voting matrix. Table 1 shows a Three Alternatives and 11 Voters
voting matrix with 13 voters and three
candidates. Each entry in the matrix gives a b c
the total number of votes obtained by the
a 8 6
row candidate over the column candidate. b 5 — 11
For example, the comparison a versus b c 7 2 —
yields eight votes in favor of a, five in
favor of b. The comparison b versus c 1. All possible opinions that do not imply a con-
tradiction reduce to an indication of the order of
yields eleven votes for b and two for c. merit that one judges to exist among the candi-
The comparison a versus c yields six for a dates . . . . therefore for n candidates one would
and seven for c. Notice that this vote have n{n — 1) . . . 2 possible opinions.
entails the cyclical majority a ~>b,b > c, 2. Each voter having thus given his or her
c>a. opinion by indicating the candidates' order of
worth, if one compares them two by two, one
Given a series of pairwise votes as in will have in each opinion n(n — l)/2 proposi-
Table 1, the problem posed by Condorcet tions to consider separately. Taking the number
is to find the ranking of all the candidates of times that each is contained in the opinion of
one of the q voters, one will have the number of
that is most likely to be correct, that is, voices who are for each proposition.
the ranking that would have produced the 3. One forms an opinion from those n(n —
observed votes with highest probability. l)/2 propositions that agree with the most
At first blush it might appear that this voices. If this opinion is among the n(n — 1). . .
problem is insoluble unless the compe- 2 possible opinions, one regards as elected the
subject to whom this opinion accords the pref-
tency levels of the voters are known. This erence. If this opinion is among the 2" (n ~ 1>/2 —
difficulty can be circumvented if one n(n — 1 ) . . . 2 impossible opinions, then one suc-
makes the simplifying assumption (which cessively deletes from that impossible opinion
Condorcet did) that the voters all have the the propositions that have the least plurality, and
one adopts the opinion from those that remain.
same competence. In this case the answer (1785, 125-26)
turns out to be independent of the level of
competence, as I shall show in the next Let us apply this rule to the situation in
section. Here we shall content ourselves Table 1. First we are to choose the three
with Condorcet's statement of the solu- propositions having a majority, namely b
tion. > c with eleven votes, a > b with eight
First some terminology is required. In votes, and c > a with seven votes. Since
Condorcet's lexicon, an "opinion" is a these three propositions form a cycle,
series of pairwise comparisons on the however, we delete the proposition with
alternatives. Each pairwise comparison is smallest plurality, namely c > a. This
called a "proposition" and written a > b, leaves the combination b > c and a > b
etc. An opinion is said to be "impossible," (and implicitly a > c), which implies the
"contradictory," or "absurd" if some of ranking a > b > c.
the propositions composing it form a The rule is clear enough in the case of
cycle, such as a > b, b > c, c > a. Nor- three alternatives. However, in the
mally, each individual voter is able to general case some ambiguities arise. The
rank all of the candidates in a consistent difficulty lies in the phrase "successively
order. This is not necessarily true of the deletes" in step 3. Taken literally, it means
aggregate opinion, however, as Table 1 the following: If there are cycles in the
demonstrates.4 To break such cyclic ma- propositions selected in step 2, delete first
jorities, Condorcet proposed the follow- the proposition having the least number
ing method:5 of votes in its favor. (This proposition
1233
American Political Science Review Vol. 82
1234
Condorcet's Theory
3'. One forms an opinion from those n(n — l)/2 Consider again the voting matrix in
propositions that agree with the most voices. If Table 1. Suppose that the true ranking is
this opinion is among the n(n — 1 ) . . . 2 possible
opinions, one regards as elected the subject to abc—a first, b second, c third. A vote for
whom this opinion accords the preference. If this a over b will occur with probability 1 —
opinion is among the 2" (n ~ 1)/2 - n(n — 1). . . 2 p, whereas a vote for b over a will occur
impossible opinions, then one reverses in that with the lesser probability 1 — p. The
impossible opinion the set of propositions that
have the least combined plurality and one adopts combined probability of observing the 39
the opinion from those that remain. pairwise votes in Table 1 is therefore
It seems reasonably likely that this is Uabc) [p s (l - p)5]
what Condorcet meant to say. In any 815!
case, it is the only interpretation that is
consistent with his goal of finding the i3!
ranking that is most likely to be correct. 617! 1112!
= (131)3 [p25(1_p)14]_
1235
American Political Science Review Vol. 82
1236
Condorcet's Theory
1237
American Political Science Review Vol. 82
Therefore the posterior probability that c approximately as follows:
is best is
a: (1/2)"° + (l/2)"»(23 - 37 + 29 - 31)e
Pr(c|vote) = b: (1/2)120 + (l/2)"'(37 - 23 + 29 - 31)e
Pr(c > a|vote)Pr(c fc|vote) c: (1/2)120 + (1/2)"'(31 - 29 + 31 - 29)e.
approach that has had enormous influ- While Condorcet's solution is attrac-
ence on the modern theory of social tive, Borda's rule still has advantages that
choice. To illustrate Condorcet's new- cannot be dismissed easily. First, it is con-
found argument, consider the example in siderably simpler to compute than Con-
Table 4. If a candidate, such as c, obtains dorcet's ranking rule. Second, if one
a simple majority over every other candi- accepts Condorcet's basic premises, then
date then, Condorcet argued, that candi- the Borda winner is in fact a better esti-
date is the only reasonable choice. Indeed, mate of the best candidate provided that p
candidate a has no reasonable claim to is close to 1/2. Third, if p is not close to
being best, therefore a should be set aside. 1/2, then it is still very likely that the
So the contest is between b and c. But c Borda winner is the best candidate, even
defeats b, so the choice must come down though strictly speaking it may not be the
on the side of c. optimum estimate of the best candidate.14
More generally, Condorcet proposed If there are a large number of voters and p
that whenever a candidate obtains a sim- is not very close to 1/2, then the probabil-
ple majority over every other candidate, ity is very high that the truly best candi-
then that candidate is presumptively the date will be selected by any reasonable
"best." This decision rule is now known as choice rule (i.e., with high probability it
"Condorcet's criterion," and such a candi- will be the majority winner and the Borda
date (if it exists) is a "Condorcet win- winner at the same time). The critical case
ner" or a "majority candidate" (Fishburn is precisely when p is near 1/2, and then
1973). Borda's rule is definitely optimal. Thus we
In advancing this new approach, Con- must either jettison the whole idea of
dorcet was not abandoning his basic selecting the "best" candidate with high
premise that the object of voting is to probability or admit that Borda's rule is a
make the "correct" choice. He merely good way of estimating which candidate
shifted his ground and argued for adopt- that is.
ing a simple estimate of the best choice On the other hand, when the problem is
rather than a complex one. Furthermore, to rank a set of alternatives, Condorcet's
this "new" approach is entirely consistent rule is undoubtedly better than Borda's.
with Condorcet's earlier solution of the Furthermore, even if one does not accept
ranking problem. Indeed, if a majority the specific probability model on which
candidate exists, then it must be ranked this conclusion is based, Condorcet's rule
first in a Condorcet ranking. The reason is has an important stability property that
simple. Suppose that x is a Condorcet Borda's rule lacks, as I shall now show.
candidate but that this candidate is not
ranked on top in the Condorcet ranking.
Then it must be ranked just below some Local Stability and
other candidate y. By assumption, x Nonmanipulability
defeats y by a simple majority. Therefore,
if we switch only the order of x and y in One of the most important concepts in
the ranking, then we obtain a new rank- the theory of social choice is independ-
ing that is supported by more pairwise ence: a social decision procedure should
votes. But this contradicts the definition depend only on the voters' preferences for
of the Condorcet ranking as the one sup- the "relevant" alternatives, not on their
ported by the maximum number of pair- preferences for alternatives that are in-
wise votes. Thus Condorcet's choice rule feasible or outside the domain of dis-
is consistent with his ranking rule when- course. Independence has been given sev-
ever a majority candidate exists. eral distinct formulations (Arrow 1963;
1239
American Political Science Review Vol. 82
a b c d e f Borda Scores
a 1 _ 51 54 1 58 60 62 285
b 1 49 68 | 56 51 58 283
c | 46 32 70 66 75 289
- 1
d 42 44 30 41 64 221
e 40 48 34 59 34 215
f 38 42 25 36 66 207
Nash 1950). The one that I shall adopt ones. Consider the following example: A
here is closest in spirit to Nash's version: if legislature of 100 members is considering
the set of alternatives shrinks, then the re- three alternative decisions a, b, and c (see
duced set of alternatives should be ranked Table 5). Other alternatives might in
in the same way that they were within the theory be considered, such as d, e, and /,
larger ranking. There are at least two rea- but they are weaker than a, b, and c in the
sons why independence is desirable. First, sense that any one of them would be de-
it says that decisions cannot be manipu- feated by any one of a, b, or c.
lated by introducing extraneous alterna- The question is whether the exclusion
tives. Second, it allows sensible decisions of the nominally "inferior" alternatives d,
to be made without requiring an evalua- e, f affect the choice between a, b, and c.
tion of all possible decisions. In practice, Put differently, can an inferior alternative
society cannot consider all possible be introduced strategically in order to
choices in a situation. It cannot, or at least manipulate the choice between the
usually does not, even consider all of the "superior" alternatives? This is surely a
best choices. In choosing a candidate for situation that we would like to avoid. Un-
public office, for example, it is impractical fortunately Borda's rule suffers from this
to consider all eligible citizens. In con- defect. To see this, observe that the total
sidering a piece of legislation, it is impos- Borda score of each alternative is simply
sible to hold votes on all possible amend- the sum of the row entries corresponding
ments. Independence assures that a deci- to that alternative (e.g., the right-hand
sion made from a limited agenda of alter- column of Table 5). These scores imply
natives is valid relative to that agenda. the ranking cabdef. Now consider just the
Enlarging the agenda may introduce bet- three alternatives «, b, and c. The cor-
ter possibilities, but it should not cause us responding voting matrix is the one en-
to revise the relative ranking of the old closed by dashed lines. The Borda scores
possibilities. for this three-alternative situation are 117
Unfortunately, it is impossible to design for b, 105 for a, and 78 for c. Hence when
any reasonable social decision rule with just the top three alternatives are con-
this property. I claim, however, that sidered, they would be ordered bac,
independence can be satisfied if we restrict which is the reverse of how the electorate
ourselves to agendas that are sufficiently would have ordered them in the context
"connected." By a connected agenda I of all alternatives. Introducing the inferior
mean, roughly speaking, a subset of close- alternatives changes the outcome from b
ly related alternatives rather than an arbi- toe.
trary collection of unrelated or extreme I claim that Condorcet's rule is immune
1240
Condorcet's Theory
to this type of instability. To compute the tion that abcdfe is supported by the maxi-
Condorcet solution observe first that a mum number of pairwise votes. A similar
has a simple majority over every other argument applies if the last two alterna-
alternative, hence (as was demonstrated tives are dropped, the last four, or indeed
in the previous section) Condorcet's rule any bottom segment of the ranking. Simi-
must rank a first. Since each of the alter- larly, the ordering of the bottom part of
natives a, b, and c defeats each of the the list will be preserved if we remove the
alternatives d, e, and / by a majority, it top alternative, the top two alternatives,
also follows that the former three alter- or any top segment of the ranking.
natives must be ranked above the latter In general, a Condorcet ranking has the
three. For if this were not the case, then property that the ordering of the alter-
some alternative would be ranked im- natives does not change whenever we
mediately below another alternative that restrict attention to an interval of the full
it defeats, which is impossible in a Con- ordering.15 This property will be called
dorcet ranking. To maximize agreement local stability. Intervals tend to consist of
with the pairwise voting data, a must be alternatives that are relatively closely
ranked first, b second, and c third. Posi- related. For example, an interval might
tions four, five, and six in the ranking will consist of a set of candidates that occupy
be occupied by d, e, and / in some order. a certain segment of the political spectrum
To determine which order, it suffices to (e.g., "centrist"). Or it might consist of a
apply Condorcet's criterion to these three series of minor modifications or amend-
alternatives alone; that is, they should be ments to a proposed piece of legislation.
ordered so as to agree with the maximum I shall show that Condorcet's method is
number of pairwise votes that they the only plausible ranking procedure that
receive among themselves. The reader is locally stable. To do so we must recall
may verify that the ordering that accom- several properties that are standard in the
plishes this objective is dfe. Hence the social choice literature. Let F be a social
total ordering implied by Condorcet's rule ranking rule, that is, a rule for strictly
is abcdfe. This is quite different from the ordering any set of alternatives given the
Borda ordering. preference data of the voters. We shall
I have worked out the details of this assume that the input data to F is given in
example to show that Condorcet's rule is the form of a voting matrix: the number
not necessarily difficult to compute even of votes for and against each pair of alter-
when a sizable number of alternatives is natives. All voters are weighted equally
involved (though it is certainly more diffi- because only the number of votes counts,
cult than Borda's method). It also illus- not the identities of the individuals who
trates why Condorcet's method satisfies a cast the votes. This property is known in
certain form of independence. For exam- the literature as anonymity. F is neutral if
ple, if the bottom three alternatives are it is not "biased" with respect to one or
dropped from consideration, then Con- another alternative: if the names of the
dorcet's rule (unlike Borda's) must pre- alternatives are permuted, then their
serve the ordering abc. Indeed, if it did ranking is similarly permuted. Taken
not, then the interval abc could be re- together, anonymity and neutrality re-
arranged into a new ordering that is sup- quire that F sometimes yields ties.16 A
ported by more pairwise votes. But then third standard condition in the social
abc could also be rearranged within the choice literature is that F be unanimous: if
full ordering abcdfe to obtain a new all voters cast their votes the same way on
ordering that is supported by more pair- every pair, then the ordering of the candi-
wise votes. This contradicts the assump- dates agrees with all of the votes.
1241
American Political Science Review Vol. 82
1242
Condorcet's Theory
Condorcet's method is a rational way of 10. Condorcet arrived at this answer intuitively,
aggregating individual choices into a col- not by an explicit application of Bayes' rule.
11. This "minimax" rule was proposed in modern
lective preference ordering. times by Kramer (1977).
12. Borda himself gave this alternative statement
of his method.
13. This idea was conjectured independently by
Grofman (1981). The distinction between the major-
Notes ity alternative and the alternative most likely to be
best has also been discussed by Baker (1975) and
This work was supported by the Office of Naval Urken and Traflet (1984).
Research under contract N00014-86-K-0586 at the 14. The following argument was suggested by
University of Maryland. The author is indebted to Bernard Grofman.
Bernard Grofman and William Riker for very help- 15. This property of Condorcet's method (which
ful comments on the manuscript. is known in the literature as the "median proce-
1. Condorcet did not explicitly say that he was dure") was noted by Jacquet-Lagreze (1969). For
addressing Rousseau's problem, but Rousseau's other results on the median procedure see Michaud
ideas would have been much discussed in Condor- and Marcotorchino (1978), Barthelemy and Mon-
cet's circles. Baker (1975, 229) also suggests that jardet (1981), Michaud (1985), and Barthelemy and
Condorcet drew his inspiration from Rousseau. McMorris (1986).
Grofman and Feld (1988) explore the connections 16. For example, if there are an even number of
between Condorcet's and Rousseau's ideas in greater voters and every pairwise vote is a tie, then anony-
depth. mity and neutrality imply that all linear orderings of
2. Poisson (1837) studied a similar set of ques- the alternatives are equally valid. Thus F must some-
tions using probabilistic methods. See also Gelfand times take on multiple values.
and Solomon (1973) for recent work in the Poisson 17. We require furthermore that if another rank-
tradition. ing is approved by only one of the two groups, then
3. A related problem is to rank teams or players it is not tied with a ranking that is approved by both
based on their win-loss records (as in baseball, ten- of them.
nis, or chess). Condorcet's model can be used to 18. Let F satisfy the five conditions of the theo-
estimate the most likely true ordering of the players rem. When there are exactly two alternatives, it
(see Batchelder and Bershad 1979; Good 1955; Jech follows that F is simple majority rule (Young 1974).
1983). Now consider the case of more than two alterna-
4. The following preference data would yield the tives. Let R be the social order. Local stability im-
pairwise votes in Table 1: six voters have preference plies that whenever two alternatives are ranked one
abc, five voters have preference bca, and two voters immediately above another in R, then their ordering
have preference cab. It bears emphasizing, however, remains invariant when F is restricted to those two
that Condorcet did not think of the voters' opinions alternatives alone. Since F is simple majority rule on
as expressing their personal preferences. Rather, they every two alternatives, it follows that in the social
represent the voters' best judgments as to which can- order R every alternative must defeat (or tie) the
didate in each pair is the better one. alternative ranked immediately below it. Further-
5. In this translation I have modified the mathe- more, if the two alternatives are tied, then they may
matical notation slightly to conform with modern be switched to obtain a social ordering tied with R. It
conventions. can be shown by induction on the number of alter-
6. Michaud (1985) comes to a similar conclu- natives that this property, together with neutrality,
sion. unanimity, and reinforcement implies that F is Con-
7. I exclude the case p = 1, because this would dorcet's ranking rule (Young and Levenglick 1978).
imply that all voters agree with probability one. The
case where the voters have different a priori proba-
bilities of being right is treated in Young 1986 (see References
also Nitzan and Paroush [1982] and Shapley and
Grofman [1984]). Arrow, Kenneth J. 1963. Social Choice and Indi-
8. For reasons that remain obscure, Borda's vidual Values. 2d ed. New York: John Wiley.
method was jettisoned by Napoleon after his own Baker, Keith M. 1975. Condorcet. Chicago: Univer-
election to membership (Mascart 1919). sity of Chicago Press.
9. In a footnote, Condorcet adds that his own Barthelemy, J. P., and F. R. McMorris. 1986. "The
treatise "had been printed in its entirety before I had Median Procedure for n-Trees." Journal of Clas-
any acquaintance with this method [of Borda], apart sification 3:329-34.
from the fact that I had heard several people speak Barthelemy, J. P., and Bernard Monjardet. 1981.
of it" (1785, clxxix). "The Median Procedure in Cluster Analysis and
1243
American Political Science Review Vol. 82
1244